The univariate polynomial B-spline is a real-valued function on a closed interval defined with the aid of a sequence of non decreasing points, the so-called knot sequence with . B-splines are defined using divided differences. Given an integer and a non-decreasing knot sequence , the th B-spline of order k for the knot sequence is denoted by and is given by [14]:
where is the piecewise truncated power function
of order k [14]:
The set of B-splines forms a basis that spans a vectorial space . Any polynomial spline function of order k in can be formulated as a linear combination of the B-splines basis functions :
The coefficients can be selected so as to enforce specific conditions on the spline function. For instance by enforcing that the spline function agrees with a given data set values () we get the spline interpolation conditions:
This is a linear system of equations in the coefficients 's. The system is invertible provided that:
Similar linear system can be setup for other conditions such as least squares approximation.
The imposed conditions can be generalized through the use of some bounded nonlinear functional and the nonlinear least squares objective as follows:
where the 's are the observed functional values and a properly selected subspace with basis function . Choosing an appropriate subspace is crucial. The degree and general qualitative properties such as the smoothness of the function can be determined based on expectations about the solution in a straightforward fashion. However, specifying the appropriate knot sequence () is a more intricate and difficult problem. No algorithmic solution to the knot placement problem exists in general for the nonlinear problem. Problem specific strategies can be formulated based on heuristics and prior knowledge of the physical laws involved (e.g. [57]).
Once the proper subspace is specified the solution of (2.76) reduces to the determination of appropriate coefficients that will minimize the least squares objective. This is a discrete parameters extraction problem which can be solved by standard nonlinear optimization techniques.
Spline models of the form of (2.73) are non-parameteric models. A non-parameteric regression model assumes general qualitative properties of the underlying function which are satisfied if it is selected from a collection of functions such as above. This allows greater flexibility in the possible form of the function without the inherent assumptions necessary in using parametric models. Nonparametric models rely more heavily on data and are unbiased by a priori information.
Given a set of breakpoints for the spline function , a recommended knot sequence is given for degree by:
This knot sequence ensures the continuity of and its first derivatives in .